U.S. patent application number 13/327527 was filed with the patent office on 2012-07-19 for systems and methods for wind forecasting and grid management.
This patent application is currently assigned to Vaisala, Inc.. Invention is credited to Richard Pyle, Jianmin Shao, Nicholas Wilson.
Application Number | 20120185414 13/327527 |
Document ID | / |
Family ID | 46491530 |
Filed Date | 2012-07-19 |
United States Patent
Application |
20120185414 |
Kind Code |
A1 |
Pyle; Richard ; et
al. |
July 19, 2012 |
SYSTEMS AND METHODS FOR WIND FORECASTING AND GRID MANAGEMENT
Abstract
In one embodiment, a wind power ramp event nowcasting system
includes a wind condition analyzer for detecting a wind power ramp
signal; a sensor array, situated in an area relative to a wind
farm, the sensor array providing data to the wind condition
analyzer; a mesoscale numerical model; a neural network pattern
recognizer; and a statistical forecast model, wherein the
statistical model receives input from the wind condition analyzer,
the mesoscale numerical model, and the neural network pattern
recognizer; and the statistical forecast model outputs a time and
duration for the wind power ramp event (WPRE) for the wind
farm.
Inventors: |
Pyle; Richard; (Longmont,
CO) ; Wilson; Nicholas; (Louisville, CO) ;
Shao; Jianmin; (Birmingham, GB) |
Assignee: |
Vaisala, Inc.
Louisville
CO
|
Family ID: |
46491530 |
Appl. No.: |
13/327527 |
Filed: |
December 15, 2011 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
61423507 |
Dec 15, 2010 |
|
|
|
Current U.S.
Class: |
706/11 ; 703/2;
706/21 |
Current CPC
Class: |
Y02A 90/14 20180101;
G01W 2203/00 20130101; F05B 2260/8211 20130101; G01W 1/10
20130101 |
Class at
Publication: |
706/11 ; 703/2;
706/21 |
International
Class: |
G06F 15/18 20060101
G06F015/18; G06F 17/00 20060101 G06F017/00; G06F 17/10 20060101
G06F017/10 |
Claims
1. A wind power ramp event nowcasting system comprising: a wind
condition analyzer for detecting a wind power ramp signal; a sensor
array, situated in an area relative to a wind farm, the sensor
array providing data to the wind condition analyzer; a physical
numerical model; a neural network pattern recognizer; and a
statistical forecast model, wherein the statistical model receives
input from the wind condition analyzer, the physical numerical
model, and the neural network pattern recognizer; and the
statistical forecast model outputs a time and duration for the wind
power ramp event (WPRE) for the wind farm.
2. The system of claim 1 wherein the wind condition analyzer
includes: a surface observation analyzer; and a vertical
atmospheric analyzer.
3. The system of claim 2 wherein the surface observation analyzer
detects a significant change in wind speed, strong vertical and
horizontal wind shears, a pressure drop or surge at the surface, a
temperature increase or decrease, and shifts in atmospheric
stability; the data is provided from an Atmospheric Observation
Network (AON), which is part of the sensor array; and the surface
observation analyzer includes a module for detecting and
calculating winds, pressure, temperature, and humidity.
4. The system of claim 2 wherein the vertical atmospheric analyzer
provides vertical profiles of horizontal wind speed and direction
and boundary and mixing layer heights and atmospheric
instability.
5. The system of claim 1 wherein the neural network pattern
recognizer is trained by providing it teaching patterns.
6. The system of claim 5 wherein the neural network pattern
recognizer changes according to a learning rule.
7. The system of claim 5 wherein the teaching patterns are upwind
meteorological variables.
8. The system of claim 7 wherein the upwind meteorological
variables are wind speed, wind direction, pressure, temperature,
and humidity.
9. The system of claim 5 wherein the teaching patterns are data
sets which involve wind power ramp events (WPREs).
10. The system of claim 1, further comprising: a radar analyzer,
which provides input to the statistical forecast model.
11. The system of claim 1, further comprising: a Lagrangian Scalar
Integration analyzer, which provides input to the statistical
forecast model.
12. The system of claim 1 wherein the physical numerical model is a
mesoscale numerical model.
13. A wind forecasting system comprising: a wind condition analyzer
for detecting a wind event signal; a sensor array situated in an
area, the sensor array providing data to the wind condition
analyzer; a mesoscale numerical model; a neural network pattern
recognizer; and a statistical forecast model, wherein the
statistical model receives input from the wind condition analyzer,
the mesoscale numerical model, and the neural network pattern
recognizer, and the statistical forecast model outputs a wind event
for the area.
14. The system of claim 13 wherein the neural network pattern
recognizer is trained by providing it teaching patterns, and the
neural network pattern recognizer changes according to a learning
rule.
15. The system of claim 13 wherein the teaching patterns are upwind
meteorological variables and the upwind meteorological variables
are wind speed, wind direction, pressure, temperature, and
humidity.
16. The system of claim 13 wherein the teaching patterns are data
sets which involve prior occurrences of wind events similar to the
wind event, and the prior occurrences of wind events are selected
from the group consisting of events that caused a wind power ramp
event (WPRE), tornados, thunderstorms, and storms with damaging
winds.
17. A graphical user interface system for managing a wind farm, the
system comprising: (a) a wind power ramp event (WPRE) prediction
window, showing predicted wind and WPRE events; (b) an Atmospheric
Observation Network (AON) monitoring window, showing the AON
network and the wind farm; (c) a ramp event message window showing
ramp event alerts; (d) a ramp event classification screen,
providing for classification of ramp events; and (e) a history
window, providing a history of past events and wind generation
statistics.
18. A method of providing a wind event forecast, the method
comprising: (a) detecting a footprint of a wind event that will
occur in the future for an area of interest with a first module;
(b) determining a duration and intensity of the wind event with a
second module; and (c) providing the wind event forecast.
19. The method of claim 18, further comprising: (d) providing
sensor data from a sensor array to the first module that is upwind
of the area of interest, wherein the sensor data is used to detect
the footprint; (e) providing the sensor data from the sensor array
to the second module, wherein the sensor data is used in the
determining of (d), wherein the first module includes a surface
observation analyzer and a vertical atmospheric analyzer and the
first module detects fronts as part of detecting the footprint, the
fronts are marked by changes in temperature, moisture, wind speed
and direction, atmospheric pressure, and a change in the
precipitation pattern; the first module detects mesoscale features
as part of detecting the footprint, wherein the mesoscale features
are marked by an increase in cumuliform clouds and rain showers;
and the first module detects dry lines, outflow boundaries/squall
lines, lee troughs, and sea/lake breezes as part of detecting the
footprint.
20. The method of claim 19 wherein the second module includes a
neural network, a mesoscale numerical model, and a physical
numerical model.
21. The method of claim 19, further comprising (d) training the
neural network by providing weather data, wherein the weather data
is data sets which involve WPREs only.
Description
CROSS-REFERENCE TO RELATED APPLICATION
[0001] This application claims the benefit under 35 USC 119(e) of
U.S. Provisional Patent Application No. 61/423,507 filed Dec. 15,
2010, which provisional application is hereby incorporated by
reference to the same extent as though fully disclosed herein.
BACKGROUND OF THE INVENTION
[0002] Wind energy, generated by winds, is fast becoming an
important renewable energy resource, as well as a stable,
profitable, and low-risk investment worldwide. Based on available
figures from 11 of the top 15 countries representing over 80% of
the world market, World Wind Energy Association (WWEA) recorded
5374 megawatts (MW) of new installed capacity in the first quarter
of 2009, equaling an increase of 23% compared with last year in the
same countries. WWEA keeps its previous prevision of a total
installed capacity of 152,000 MW worldwide by the end of 2009,
which will mean a new record of over 30,000 MW of newly installed
capacity within one year. This represents a market growth of 25%
compared with last year.
[0003] By comparison with other renewable energy sources, wind
energy is associated with one of the lowest costs of electricity
production and the largest resource availability. Because of its
technical maturity and efficiency, wind energy has been widely
regarded as one of the most prominent energy resources in the
future. The intermittent and volatile nature of winds poses several
problems to wind power generation. One of the largest problems, as
compared to conventionally generated electricity, is that wind
power is dependent on volatile wind. Wind volatility occurs on all
time scales, from milliseconds to days, and wind volatility impacts
both individual turbine control and the integration of wind power
into an electrical grid network. Wind volatility poses various
challenges to the wind energy industry.
[0004] Many of the challenges posed by wind volatility are
connected to the requirements for forecasting wind energy
production with a certain degree of accuracy. Increases in the
accuracy of wind energy forecasts reduce the requirement for backup
energy, resulting in increased power grid reliability, as well as
significant monetary savings. Accurate prediction of wind power is
crucial, both for technical and financial reasons. For example,
wind farm operators need the prediction of wind power to be
accurate to avoid being penalized by the system operator for not
generating power as predicted.
[0005] The development of forecasting technologies is needed to
address these challenges. An immediate requirement is for the
development of improved short-term forecasting methods, which are
necessary for transmission scheduling and resource allocation. The
short-term forecasting methods need to address areas of interest
with a greater degree of accuracy and resolution. Of central
concern are short term wind energy forecasts in the range of 0-3
hours, for example. This lead time is typical of the time necessary
for transmission scheduling and the dispatching of resources to
keep the power supply in the grid in line with power demands.
SUMMARY
[0006] In one embodiment, a wind power ramp event nowcasting system
includes a wind condition analyzer for detecting a wind power ramp
signal; a sensor array, situated in an area relative to a wind
farm, the sensor array providing data to the wind condition
analyzer; a physical numerical model; a neural network pattern
recognizer; and a statistical forecast model, wherein the
statistical model receives input from the wind condition analyzer,
the physical numerical model, and the neural network pattern
recognizer; and the statistical forecast model outputs a time and
duration for the Wind Power Ramp Event (WPRE) for the wind farm.
Optionally, the wind condition analyzer includes a surface
observation analyzer and a vertical atmospheric analyzer. In one
alternative, the surface observation analyzer detects a significant
change in wind speed. Alternatively, the surface observation
analyzer detects strong vertical and horizontal wind shears. In
another alternative, the surface observation analyzer detects a
pressure drop or surge at the surface. Alternatively, the surface
observation analyzer detects a temperature increase or decrease. In
one alternative, the physical numerical model is a mesoscale
numerical model. In yet another alternative, the surface
observation analyzer detects shifts in atmospheric stability.
Alternatively, the data is provided from an Atmospheric Observation
Network (AON) which is part of the sensor array. In yet another
alternative, the surface observation analyzer includes a module for
detecting and calculating winds, pressure, temperature, and
humidity. In yet another alternative, the vertical atmospheric
analyzer provides vertical profiles of horizontal wind speed and
direction. Optionally, the vertical atmospheric analyzer provides
boundary and mixing layer heights and atmospheric instability.
Optionally, the neural network pattern recognizer is trained by
providing it teaching patterns. In another alternative, the neural
network pattern recognizer changes according to a learning rule.
Optionally, the teaching patterns are up-wind meteorological
variables. Optionally, the up-wind meteorological variables are
wind speed, wind direction, pressure, temperature, and humidity. In
another alternative, the teaching patterns are data sets which
involve WPREs. In another alternative, the system further includes
a radar analyzer, which provides input to the statistical forecast
model. In another alternative, a Lagrangian Scalar Integration
analyzer provides input to the statistical forecast model.
[0007] In one embodiment, a wind forecasting system includes a wind
condition analyzer for detecting a wind event signal; a sensor
array, situated in an area, the sensor array providing data to the
wind condition analyzer; a mesoscale numerical model; a neural
network pattern recognizer; and a statistical forecast model,
wherein the statistical forecast model receives input from the wind
condition analyzer, the mesoscale numerical model, and the neural
network pattern recognizer, and the statistical forecast model
outputs a wind event for the area. Alternatively, the neural
network pattern recognizer is trained by providing it teaching
patterns. Alternatively, the neural network pattern recognizer
changes according to a learning rule. In another alternative, the
teaching patterns are up-wind meteorological variables. Optionally,
the up-wind meteorological variables are wind speed, wind
direction, pressure, temperature, and humidity. In another
alternative, the teaching patterns are data sets which involve
prior occurrences of wind events similar to the wind event.
Alternatively, the prior occurrences of wind events are events that
cause a WPRE. In another alternative, prior occurrences of wind
events are tornados. Alternatively, prior occurrences of wind
events are thunderstorms. Optionally, prior occurrences of wind
events are storms with damaging winds.
[0008] In one embodiment, a graphical user interface system for
managing a wind farm includes a WPRE event prediction window,
showing predicted wind and WPRE events; and an Atmospheric
Observation Network (AON) monitoring window, showing the AON
network and the wind farm. In one alternative, the system includes
a ramp event message window showing ramp event alerts. In one
alternative, the system includes a ramp event classification
screen, providing for classification of ramp events. In one
alternative, the system includes a history window, providing a
history of past events and wind generation statistics.
[0009] In one embodiment, a method of providing a comprehensive
wind forecast includes providing a first segment of the forecast,
from time t0 to t1, based on a nowcast, the nowcast created by a
statistical model receiving input from a wind condition analyzer, a
mesoscale numerical model, and a neural network pattern recognizer;
providing a second segment of the forecast, from time t1 to t2,
based on a physical model; and providing a third segment based on
National Weather Service forecasting.
[0010] In one embodiment, a method of providing a wind event
forecast includes detecting a footprint of a wind event for an area
of interest with a first module; determining a duration and
intensity of the wind event with a second module; and providing the
wind event forecast. In one alternative, the method further
includes providing sensor data from a sensor array to the first
module that is upwind of the area of interest, wherein the sensor
data is used to detect the footprint. In another embodiment the
method further includes providing the sensor data from the sensor
array to the second module, wherein the sensor data is used in the
determining. Alternatively, the first module includes a surface
observation analyzer. Optionally, the first module includes a
vertical atmospheric analyzer. Optionally, the first module detects
fronts as part of detecting the footprint. In one option, the
fronts are marked by changes in temperature, moisture, wind speed
and direction, atmospheric pressure, and a change in the
precipitation pattern. Optionally, the first module detects
mesoscale features as part of detecting the footprint. In one
alternative, the mesoscale features are marked by an increase in
cumuliform clouds and rain showers. In another alternative, the
first module detects dry lines as part of detecting the footprint.
Alternatively, the first module detects outflow boundaries/squall
lines as part of detecting the footprint. In another alternative,
the first module detects lee troughs as part of detecting the
footprint. Alternatively, the first module detects sea/lake breezes
as part of detecting the footprint. In another alternative, second
module includes a neural network. Optionally, second module
includes a mesoscale numerical model. In another alternative,
second module includes a physical numerical model. In yet another
alternative, the second module includes a neural network and
mesoscale numerical model. In one alternative, the method further
includes training the neural network by providing weather data. In
another alternative, the weather data is data sets which involve
WPREs only. Optionally, a statistical forecast model performs the
providing. Optionally, the statistical forecast model uses AR
regression. In one alternative, the footprint is fine spatial and
temporal features of conditions that correlate to the wind event.
In one alternative, the method further includes, providing weather
data from public weather sources to the first module, wherein the
public weather data is used in the predicting. In another
alternative, the method further includes providing weather data
from public weather sources to the second module, wherein the
public weather data is used in the determining.
BRIEF DESCRIPTION OF THE DRAWINGS
[0011] FIG. 1 shows an embodiment of a wind forecasting system;
[0012] FIG. 2 shows an embodiment of a wind forecasting method;
[0013] FIG. 3 shows an exemplary graph of wind conditions vs. time
depicting volatility;
[0014] FIG. 4 shows a number of plots of power provided from
various sources and a number of load obligations;
[0015] FIG. 5 shows an exemplary graph of wind speed vs. wind
power;
[0016] FIG. 6 shows an exemplary Artificial Neural Network
(ANN);
[0017] FIG. 7 shows an exemplary graph of power generation from a
variety of sources;
[0018] FIG. 8 shows locations of Atmospheric Observation Network
(AON) sites in an exemplary system;
[0019] FIG. 9 shows one embodiment of a system for predicting
WPREs;
[0020] FIG. 10 shows an example of the output from a Vertical
Atmospheric Analyzer (VAA);
[0021] FIG. 11 shows an embodiment of an interface accessed by the
user in a WPRE prediction system;
[0022] FIG. 12 shows an embodiment of a ramp event summary
table;
[0023] FIG. 13 shows a window accessible through the interface of
FIG. 11;
[0024] FIG. 14 shows one feature of the ramp window of FIG. 11;
[0025] FIG. 15 shows another feature of the ramp window of FIG.
11;
[0026] FIG. 16a-e show embodiments of a methods of detecting wind
events;
[0027] FIG. 17 shows an exemplary architecture for one embodiment
of a system for detecting wind events; and
[0028] FIG. 18 shows an embodiment of a ramp window.
DETAILED DESCRIPTION
[0029] In order to understand embodiments of a wind forecasting
system and method, aspects of wind forecasting must be examined. An
example of a useful forecast of wind informs a user of a location,
a time at which the forecasted event will occur, and the
characteristics of the forecasted event--basic what, when, and
where. In various embodiments of wind forecasting systems described
herein, the what, when, and where are described in relation to wind
farms and the characteristics of wind that affect wind farms, such
as ramp events. The characteristics of wind events or forecasted
events that affect wind farms can equally affect other areas of
interest as well, such as airports, crop regions, areas vulnerable
to tornados, scientific research area (where experiments can be
disturbed), etc.
1. Short-Term Wind Forecasts
[0030] In one embodiment, short-term wind forecasts can be targeted
for prediction. The development of improved short-term forecasting
methods is necessary for transmission scheduling and resource
allocation. Short-term forecasting can be referred to as
nowcasting. Generally, nowcasting can be a form of very short-range
weather forecasting, covering only a very specific geographic area.
For example, a nowcast can be defined as a forecast for the coming
12-hour period, based on very detailed observational data. More
specifically, a nowcasting can be a detailed description of current
weather conditions, from which one can extrapolate (project) the
weather conditions for the following two hours. Of central focus
for wind energy systems are wind energy forecasts on the 0-3 hour
time horizon. A certain amount of lead time of 0-3 hours is
necessary for transmission scheduling and the dispatching of
resources.
[0031] In this short term forecast, for example, in the 0-3 hour
time horizon, a significant challenge is the prediction of wind
power ramp events. Wind power ramp events can be caused, in most
cases, by an abrupt change in wind speed. They result in
significant changes of the schedule of power generation over a
short period and can cost wind power operators significantly. If
the utility company is caught off guard and cannot accurately and
reliably predict wind power ramp events, significant energy
management issues arise. On the other hand, the company can work to
reallocate or balance the energy grid if these incidents are
properly predicted. In well-developed financial markets like the
United States, accurate forecasts of wind power output can be
valuable in the derivative energy market. Therefore, an accurate
nowcasting (0-3 hours ahead) for example systems and methods for
wind power ramp events are provided. Systems and methods described
herein can be key components of a decision support system that
helps wind power operators to minimize the adverse impacts of the
ramp events on power generation and schedule.
[0032] In one aspect, a decision support system and method are
provided that can be utilized by electric utilities and balancing
authorities to improve reliability and financial performance by
improving the short-term (0-to-3 hour, for example) forecast of
wind energy generation and the timing, magnitude, and
rate-of-change of ramp events to most effectively dispatch and/or
curtail other power generators.
[0033] 1.1 Sensor Array and Model-Based Prediction
[0034] In one embodiment, a wind forecasting system is used to
predict a wind event in an area of interest 100 as illustrated in
FIG. 1. The area of interest 100 may vary in size. A plurality of
sensors 110 detects a plurality of conditions. The plurality of
sensors 110 can be located remotely from the area of interest 100.
The exact location of the plurality of sensors may vary; however,
in the embodiment of FIG. 1, the plurality of sensors 110 is
located approximately a distance d from the area of interest 100,
where d is based on the average wind speed (w) of the winds in the
direction 120 multiplied by the time (t) in advance the forecast is
desired (d=w*t). For example, if the typical winds in the direction
120 are 40 miles per hour and a one hour advance forecast is
desired, then the sensors 110 are located approximately 40 miles
from area 100. Alternatively, sensors are located at varying
distances, some close to the area of interest 100 and some farther
away. [Insert limit to sensor distances.]
[0035] Sensors 110 detect a plurality of conditions. Conditions
include, but are not limited to, wind conditions, humidity,
electrical activity, temperature, etc. More information on detailed
characteristics of weather conditions is described below. The
plurality of conditions is processed by footprint wind event
prediction module 130. The processes of the footprint wind event
prediction module 130 are described in more detail below, in
relation to weather conditions, algorithms, and models. In this
embodiment, the footprint wind event prediction module 130 predicts
when the wind event will occur. The area of interest 100 is where
the wind event will occur. In order to predict the duration of the
wind event, model-based wind prediction module 140 receives data
from the plurality of sensors 110 and weather condition and
forecast sources 150. Based on this input, model-based wind
prediction module 140 predicts the intensity and duration of the
wind event. The processes of the model-based wind prediction module
140 are described in more detail below, in relation to weather
conditions, algorithms, and models. In many alternatives,
additional sources of data are provided to the modules by National
Weather Service Information and turbine sensor data.
[0036] In one embodiment of a method of wind forecasting, as shown
in FIG. 2, in step 210 a plurality of sensors measure conditions.
The sensors are located remote from the area of interest, by a
distance d, similar to that described above in relation to FIG. 1.
In step 220, data is provided to a footprint wind event prediction
module. In step 230, the footprint wind event prediction module
predicts the time of occurrence of a wind event. The processes of
the footprint wind event prediction module are described in more
detail below, in relation to weather conditions, algorithms, and
models. In step 240, the plurality of sensors and other weather
condition and forecast sources provide data to a model-based wind
prediction module. In step 250, the model-based wind prediction
module predicts the intensity and duration of the wind event.
[0037] In the embodiments described in relation to FIGS. 1 and 2,
the systems and methods predict conditions periodically, or
essentially constantly (accounting for the time required for the
modules to process the data and make a prediction), or any measure
in between. In one alternative to the system and method described
above in relation to FIGS. 1 and 2, the model-based wind event
prediction module does not receive data from the plurality of
sensors. Below are described alternative embodiments to those
described in relation to FIGS. 1 and 2. Although the embodiments
are described in terms of discrete embodiments or sub portions of
discrete embodiments, any of the aspects may be substituted in part
or in whole into any of the other systems and methods described
herein. Sensors used for the plurality of sensors 110 are described
below. The wind prediction modules described above are implemented
in general purpose computing hardware and software; however,
alternatively, parts or all of the system are implemented in more
hardware-based systems as well, such as field programmable gate
arrays (FPGAs) or other more hardwired integrated circuits.
[0038] In one embodiment, shown in FIG. 16a, given an area of
interest, a first module detects the footprint of a wind event that
will affect that area of interest in step 1605. A second module
determines the duration and intensity of the wind event in step
1607. In step 1610 the wind event is forecast. In one alternative,
shown in FIG. 16b, sensor data is provided by a sensor array to the
first module that is upwind of the area of interest in step 1612.
Upwind, in this case, means that if the wind is coming from the
North, then the sensor array is further north than the area of
interest. In another alternative, the sensor data is provided to
the second module. In another alternative, the sensor data is
provided to both modules.
[0039] The first module may include a number of different footprint
detection systems. One option is a surface observation analyzer.
Another option is a vertical atmospheric analyzer. Another option
is both a surface observation analyzer and a vertical atmospheric
analyzer. The first module uses Fronts (marked by changes in
temperature, moisture, wind speed and direction, atmospheric
pressure, and often a change in the precipitation pattern),
Mesoscale Features, such as shear line (marked by an increase in
cumuliform clouds, often including towering cumulus and rain
showers), dry lines (marked by downsloped air from higher terrain
mixed to the surface during daytime heating, resemblance to a cold
front although sloshing eastward during the day, and westward at
night), outflow boundaries/squall lines (marked by convection that
is strong and linear/curved, with the feature placed at the leading
edge of the significant wind shift and pressure rise), lee troughs
(marked by westerly winds aloft increase on the north side of
surface highs, areas of lowered pressure that form downwind of
north-south oriented mountain chains), sea/lake breezes (marked by
during the afternoon, hot air on land ascends and a sea/lake breeze
moves inland in the vacancy left by lowered pressures formed with
warmed air over the land and the reverse at night), and Microscale
features (may be ignored in the case of a large enough wind farm).
In one alternative, the first module receives the sensor data from
the sensor array. In another alternative, the first module receives
sensor data from the sensor array and forecast data from other
sources (for example the National Weather Service). In another
alternative, the first module does not receive data from the sensor
array, instead receiving data from remote sources (for example, the
National Weather Service). In an alternative, sensor data is
provided by individuals with home monitoring stations. Optionally,
these home monitoring stations are connected to the Internet to
provide frequent electronic updates. The first module determines
the time of occurrence of the wind event. The processes of the
first module are described in more detail below, in relation to
weather conditions, algorithms, and models. FIG. 16c shows one
embodiment of a method using a surface observation analyzer and a
vertical atmospheric analyzer. Sensor data is provided by a sensor
array to the first module that is upwind of the area of interest in
step 1612. In step 1615 a surface observation analyzer detects the
footprint of a wind event. In step 1617 a vertical atmospheric
analyzer detects the footprint of a wind event. In step 1620, if a
footprint is detected, then the method proceeds to step 1607 where
the second module detects the time and duration of the wind event.
In step 1610 a wind event is forecast.
[0040] The second module may include a number of different duration
and intensity detectors. One option is a neural network. Another
option is a mesoscale numerical model. Another option is a physical
numerical model. Another option is an AR model. Another option is
multivariate regression. Another option is Model Output Statistics.
Another option is a combination of a neural network and a mesoscale
numerical model. Another option is a combination of any number of
the above. In one embodiment, the second module and the first
module are combined into a single system. In one embodiment, the
second module and the first module share some common
components.
[0041] In an option including a neural network and mesoscale
numerical model, the mesoscale model uses mesoscale features
provided by outside forecast sources and the sensor data from the
sensor array. More detail concerning the neural network and the
mesoscale model are provided below. The mesoscale numerical model
detects and quantifies Mesoscale Features, such as shear line
(marked by an increase in cumuliform clouds, often including
towering cumulus and rain showers). Fast four-dimensional data
assimilation scheme is used to take advantage of the sensor data.
Results mesoscale numerical model is used to aid the neural
network. The neural network is trained by feeding it weather data.
In one alternative, the variables as inputs to the neural network
are mainly up-wind meteorological variables like wind speed, wind
direction, pressure, temperature, and humidity. In one alternative,
instead of training the neural network indiscriminately based on
all of the historical data, the neural network uses carefully
selected data sets which involve WPREs only. The steps that take
place at the step 1607 may be expanded as shown in FIG. 16d. In
this alternative, step 1620 a Mesoscale Model Module detects
mesoscale features. In step 1622 these features are provided to a
Neural Network with other information concerning weather
conditions. In step 1624, the Neural Network determines the time
and intensity of the wind event. In step 1626, these features are
output.
[0042] In alternative to the embodiment that includes, a first
module that detects the footprint of a wind event that will affect
that area of interest and a second module that determines the
duration and intensity of the wind event, the embodiment further
includes a third model. The third module is a Statistical Forecast
Model. In one option, the Statistical Forecast Model may use a
multivariate regression. In another option the Statistical Forecast
Model may use an AR model. In another embodiment, the Statistical
Forecast Model is a neural network and an AR. In this alternative,
the third module receives inputs from the first module and the
second module, and determines the timing, duration, and intensity
of the wind based on a synthesis of the data from the first and
second module. In the FIGS. 16a-c, step 1610 may be expanded to
include a Statistical Forecast Model as shown in FIG. 16e. The
Statistical Forecast Model receives available data in step 1628,
and provides a forecast in step 1630.
[0043] In one alternative, the wind event is wind of a certain
magnitude. In another alternative the wind event is a physical
manifestation of wind on another system. For instance, wind and the
characteristics of wind can be predicted and then the effect of the
wind on another system can be predicted. Using known
characteristics of the system and historical data concerning the
system, a model for the affect of wind on the system may be
produced. In the case of wind farms, the wind can be predicted, and
then it can be predicted how this wind will effect the wind farm.
In an alternative, the wind and the characteristics of the wind are
not predicted. Instead, the intermediate step of predicting the
wind is eliminated and simply what will happen to the physical
system of interest is predicted. For instance, instead of
predicting the wind and then predicting from the wind a WPRE, the
WPRE event is predicted directly. In this case the models are
trained and created specifically in relation to predicted the WPRE.
There is no need to specifically model the wind, only the WPRE.
This is not meant to preclude the prediction of some of the
characteristics of the wind, however, in this alternative, it is
not necessary to predict every aspect of the wind forecast. In this
alternative, in the case of neural networks, the prediction is
therefore more direct, which eliminates some inaccuracies from the
conversion of wind to WPRE.
[0044] In alternative to the embodiment that includes, a first
module that detects the footprint of a wind event that will affect
that area of interest and a second module that determines the
duration and intensity of the wind event, the second model includes
a neural network. This system is an improvement over a system that
uses a neural network or neural network in combination with other
systems, since it informs the neural network when a wind event will
occur.
[0045] Neural networks or systems that primarily rely on neural
networks for the predictions of events suffer from the disadvantage
that neural networks are good at recognizing patterns, not
predicting events. Neural networks are constantly recognizing
patterns, since that is what they are designed to do even if no
pattern exists. Therefore, the neural network is assisted by the
first module, which informs the neural network and overall system
that now an actual wind event is occurring; now it is time to
recognize the pattern and tell more about what is happening. If the
system were primarily driven by the neural network, then the neural
network would constantly recognize patterns, even if there were not
any to recognize.
[0046] By way of analogy, humans and the human mind are programmed
to recognize faces. This programming, similar to a neural network,
causes humans to recognize faces in animate items, where no face
exists (e.g. humans recognize faces very limited cues representing
gestalt of a face: :-); although this keyboard emoticon is simply
two dot, a line, and a semicircle, it is readily recognized as a
face). Faces are recognized by the human mind anywhere where basic
facial characteristics exist. Ultimately, the human mind knows that
these objects aren't actually human faces, even though some of the
characteristics match the patterns programming into the human
neural network since other analysis abilities inform the neural
network in humans designed to recognize faces. In this way, the
wind event prediction system and neural network are informed that
an event is occurring. Now is the time to recognize the
pattern.
[0047] Described below herein is more explanation of the details
and various pieces that relate to the embodiments described in
relation to FIGS. 1, 2, and 16a-e.
[0048] 1.1.1 Wind Power Ramp Event
[0049] There are a number of wind events that are of interest to
users of a wind prediction system. One such event is a Wind Power
Ramp Event.
[0050] In one embodiment, a wind forecasting system is used to
predict wind events for a wind farm. The wind forecasting system is
based on a combination of inputs including wind conditions, a
mesoscale model, and a neural network. In this embodiment wind
conditions are used to predict the occurrence of a wind event. In
one alternative, the wind event may be a Wind Power Ramp Event
(WPRE). In general terms, a WPRE is a change in the wind that will
result in a significant change in the power produced by a wind
farm.
[0051] The relationship between power output (P) from a wind
turbine and wind speed can be expressed as:
P=1/2C.sub.p.rho.AV.sup.3 (1)
[0052] where Cp is the power coefficient, .rho. is air density, A
is the rotor swept area, and V is the wind speed. The power
coefficient describes that fraction of the power in the wind that
may be converted by the turbine into mechanical work. It has a
theoretical maximum value of 0.593. Equation (1) shows that wind
energy varies as the cube of the wind speed. Therefore, any small
change in wind speed can result in a very large variation of wind
power output.
[0053] Wind is characterized in part by volatility. FIG. 3 shows
10-min averaged wind speed at Hong Kong airport. It is seen from
the figure that wind speed 310 shows many sporadic spikes, plus a
trend in general. The scale for the X axis 320 is in hours, and the
scale for the Y axis 330 is in km/h. Such character of winds
inevitably causes dramatic ups and downs in wind power output.
Under some particular situations, a rapid change in wind speed can
cause a dramatic WPRE 410 as shown in FIG. 4.
[0054] FIG. 4 shows a number of features related to power
management. The Y-axis 470 measures in megawatts (MW), and the
X-axis 480 measures in hours. The obligation of the power company
460 is shown in the top graph. This is a measure of the power that
the power company needs to provide to service its customers. The
power company interchange 450 shows the amount of power that the
power company has to buy from other sources to serve the needs of
the customers. As can be seen in the graph, as the time of day
advances to the early morning, the power needs of the customers
goes down.
[0055] The lower graph shows the coal power production 430, the
wind power production 420, and the gas power production 440. As can
be seen, the power produced from each source is modulated in order
to keep the power supply consistent. Consequences flow from the
inability of the power company to keep the power supply constant.
For instance, if not enough power is produced, then the power
company either has to buy power from another source or brownouts
will occur. If too much power is produced and it is not properly
anticipated so that it can be sold to another source, then
regulatory fines may be incurred and portions of the electrical
grid may be damaged.
[0056] A sudden change in wind speed and wind power output are the
results of changes in weather conditions. In one alternative, a
power ramp event is defined as a power loss or gain of 10-20% of
total output. Another measure of a ramp event is defined as an
absolute change of power output greater than 100 MW. The ramp
events can be categorized as ramp-up and ramp-down events. The
former are caused by: [0057] a. Frontal systems; [0058] b. Dry
lines; [0059] c. Convection (early evening), and [0060] d. Low
level jet.
[0061] The ramp-down events are related to: [0062] a. Weakening of
pressure gradient; [0063] b. Shallow, cold air mass and turbulent
mixing; [0064] c. Sudden cooling of near-surface layer and
increased stability; and [0065] d. Turbine cut-off due to high
winds (this means that the power output is sensitive to a small
change in wind at or around cut-off speed, at which a turbine shuts
down for safety reason).
[0066] While the ramp-up events may show an annual peak in late
winter to summer, the ramp-down events show no annual pattern.
Occurrence and magnitude of all ramp events are related to weather
systems which are of different spatial and temporal scales.
[0067] 1.1.2 Other Wind Events
[0068] Numerous other wind events are of interest to users.
Although much of the present disclosure is described in relation to
WPREs, the principles may be equally applied to other areas of
interest. The area of interest 100, as described in relation to
FIG. 1, clearly may be a wind farm. In alternative embodiments, the
area of interest 100 may be crops, especially those crops that may
be adversely affected by strong winds. In another alternative, the
area of interest 100 is a town or other settlement in an area with
a high risk for tornados or high winds. In another alternative, the
area of interest 100 is an airport. Any area that may be positively
or adversely affected by wind may be defined as the area of
interest. It is expected that as enough sensor networks are
deployed, the available data for predicting wind events as
disclosed herein will begin to overlap and increase, making wind
event prediction more comprehensive.
[0069] 1.1.3 Volatility
[0070] Volatility is one aspect of the characteristics of a wind
event. A volatility index (VI) is basically a measure of variation
of a variable/process from the average value (or zero) over a
certain period. A commonly used measure for volatility is the
variance. A larger variance means a volatile situation, and a
smaller variance indicates a mild/damped situation. Depending on
customer requirement, the VI, which can be normalized to 0-100 or
any range, can refer to total power output or wind speed. In case
of VI for wind speed, a separate power curve/model is needed to
convert VI for wind speed to VI for power output. As shown in FIG.
3, volatility may vary significantly over time.
[0071] Volatility is a characteristic of wind this is predicted in
some alternatives. Volatility affects how wind will affect other
physical systems. Volatility can be thought of as gustiness. For
example, a very gusty wind pattern in some cases will not produce
as much wind as a constant and consistent wind pattern, since
intermittent produces less force over time. In one alternative,
volatility is used in the prediction of the strength of a WPRE.
[0072] In one alternative, a volatility index for forecast power
output or wind speed includes forecast error or uncertainty. A
forecast error differs from volatility in that the former is
model-inherited while the latter is barely a measure for the
fluctuation of observation or forecasts around a historic mean (or
zero) over a predefined period.
[0073] The calculation of simple VI, without taking probability and
forecast uncertainty into account, can be done in the following
steps: [0074] 1. Assume the following variables: [0075] X: total
power output or wind speed; discrete [0076] n: length of history
data, discrete [0077] .mu.: mean (or zero) of X over the period of
n [0078] V: variance of X [0079] m: length of forecasts ahead,
discrete (n and .mu. should be configurable as they depend on
customer interests/requirements). [0080] 2. Calculate historic
.mu., if .mu. is not 0
[0080] .mu. = 1 n i = 1 n x i ##EQU00001## [0081] where xi is
observed X at time i. [0082] 3. Calculate historic variance
[0082] V = 1 n i = 1 n ( x i - .mu. ) 2 . ##EQU00002## [0083] 4.
Calculate VI over forecast period
[0083] VI j = ( x j - .mu. ) 2 V .times. 100 % ##EQU00003## [0084]
where xj is the forecast X at time j and j=1, 2, . . . m. [0085] 5.
If VI is based on wind speed and forecasts are wind speed, do the
following to convert wind speed volatility to power output
volatility by power model f; otherwise, skip over:
[0085] VI.sub.j'=f(VI.sub.j). [0086] This gives us a volatility
index at each forecast time/point.
[0087] In one alternative, the calculations (2-4) are based on
actual power output if customers are interested in volatility of
power output.
[0088] 1.1.4 Weather Systems
[0089] One characteristic that is used to predict the occurrence of
wind events is the occurrence of weather systems. Theses
characteristics may be considered by the prediction modules
described above and herein, based on the data provided to them. As
described above in relation to FIG. 16, the first module that
detects footprints uses many of these characteristics to predict
wind events, as do other models described below.
[0090] 1.1.4.1 Synoptic Scale Feature: Fronts
[0091] Fronts are the most common synoptic or global scale weather
features. In meteorology, they are the leading edges of air masses
with different density. When a front passes over an area, it is
marked by changes in temperature, moisture, wind speed and
direction, atmospheric pressure, and often a change in the
precipitation pattern. For example, passage of a cold front is
often accompanied by a drop in air temperature, increase in
pressure, and strong winds that are likely to cause wind ramp
events.
[0092] 1.1.4.2 Mesoscale Features
[0093] Mesoscale systems, a few hours and up to 100 km, affect wind
farms with complex interactions between the wind, moisture,
temperature, and pressure on this scale. The stability of the
atmosphere may indicate the future occurrence of many of the
phenomena. Variations on this scale determine the availability of
the wind resource and are important for scheduling and integrating
the variable generation into the national grid.
[0094] (a) Shear Line
[0095] A shear line is an area in a low pressure trough, usually in
the tropics, within which wind direction changes significantly over
a relatively short distance. The area is marked by an increase in
cumuliform clouds, often including towering cumulus and rain
showers. It may become more active with thunderstorms, and the
turbulence and circular motion of winds may assist in the formation
of a tropical storm.
[0096] (b) Dry Line
[0097] A dry line is a similar phenomenon to a frontal zone
(boundary between moist and dry air) but at a much smaller scale.
By definition, dry lines are formed in response to downsloped air
from higher terrain mixed to the surface during daytime heating. In
three dimensions, a dry line resembles a cold front, though it
normally sloshes eastward during the day, and westward at night,
caused by density differences.
[0098] (c) Outflow Boundaries/Squall Lines
[0099] Organized areas of thunderstorm activity not only reinforce
pre-existing frontal zones, but can also outrun cold fronts in a
pattern where the upper level jet splits into two streams, with the
resultant mesoscale convective system (MCS) forming at the point of
the upper level split in the wind pattern running southeast into
the warm sector parallel to low-level thickness lines. When the
convection is strong and linear/curved, the MCS is called a squall
line, with the feature placed at the leading edge of the
significant wind shift and pressure rise. Even weaker and less
organized areas of thunderstorms will lead to locally cooler air
and higher pressures, and outflow boundaries exist ahead of this
type of activity.
[0100] These features will commonly be depicted in the warm season
across the United States on surface analyses, and they lie within
surface troughs. Squall lines can cause short and sharp wind power
ramps.
[0101] (d) Lee Trough
[0102] When westerly winds aloft increase on the north side of
surface highs, areas of lowered pressure will form downwind of
north-south oriented mountain chains, leading to the formation of a
lee trough. If moisture pools along this boundary during the warm
season, it can be the focus of diurnal thunderstorms.
[0103] (e) Sea/Lake Breeze
[0104] Sea/lake breeze occurs mainly on sunny and warm days when
the land surface warms up. During the afternoon, hot air on land
ascends and a sea/lake breeze moves inland in the vacancy left by
lowered pressures formed with warmed air over the land. This
process reverses at night, leading to a land breeze and wind
acceleration offshore. If pressure gradients are large enough,
sea/lake or land breezes can cause small to moderate wind
ramps.
[0105] 1.1.4.3 Microscale Features
[0106] Microscale features are those phenomena with a timescale of
a few minutes and spatial scale of up to 1 km. They affect
individual turbines. Motions on this scale are chiefly turbulent
and irregular, and affect the choice of location for individual
turbines and the stresses on them. They can be ignored in the
prediction of aggregated power at a large scale such as for a wind
farm.
[0107] 1.1.4.4 Summary
[0108] In some embodiments, the above characteristics are used to
detect the footprint of a wind event and the duration and intensity
of the wind event. In some alternatives the surface observation
analyzer and the vertical atmospheric analyzer use these
characteristics to identify footprints.
[0109] 1.2. Wind Power Forecasts in General
[0110] There are two ways to make wind power forecasts at a large
scale: [0111] a. Forecast wind speed, and then convert wind speed
to wind power via a power curve (FIG. 5); and [0112] b. Forecast
power output directly.
[0113] In FIG. 5, the y-axis 510 has units of power output by a
wind farm in megawatts (MW). The x-axis 520 shows wind speed in
meters per second. The plotted points 530 show the conversion.
[0114] The models/methods discussed in this disclosure can be
applied in either way. However, it should be noted that the `fat`
curve or deviated spots in FIG. 5 show a great degree of
uncertainty attached to the conversion of power output from wind
speed. Factors such as volatility may affect the ability to
accurately convert wind speed to power.
[0115] In order to mitigate adverse impacts effectively from a wind
ramp event, the energy company needs to know Timing of the
occurrence; Duration of the event; and Magnitude of the ramp.
[0116] Because of volatile nature of WPREs, prediction of the
events is extremely difficult.
[0117] Prediction of wind power output can be classified as: [0118]
a. Statistical; [0119] b. Physical (or numerical); and [0120] c.
Hybrid: statistical plus physical.
[0121] All of these methodologies may be incorporated into the wind
prediction systems and methods described herein. Alternatives
described herein are simply examples of systems and methods that
may be implemented in light of this disclosure.
[0122] 1.2.1 Statistical Methods/Models
[0123] In some alternatives, statistical methods/models are used.
Wind speeds are positively and strongly correlated over a short
period of hours. This correlation is seen in that low values tend
to follow low values and high values tend to follow high values.
Furthermore, wind speeds have a property known as medium-term
memory; and its hallmark is a slowly decaying autocorrelation
function. The existence of the "memory" is important when
attempting to forecast and quantify wind-associated phenomena such
as wind power output.
[0124] As noted above, many of the statistical methods described
herein are used in embodiments of wind event prediction systems and
methods. See the above description of FIGS. 16a-e for options on
where these methodologies are incorporated to the wind event
prediction systems and methods described herein.
[0125] A collection of measurements of wind speed or observation of
wind power output over a time period is regarded as a time series,
which can be expressed as:
{x.sub.1, . . . ,x.sub.t-1,x.sub.t}.
[0126] Statistical approaches, based the time series, try to
exploit obvious or non-obvious relationships embedded in the time
series (in the past), and then apply the relationships in
extrapolation into the future.
[0127] A family of statistical methods applied in the prediction
include: [0128] a. Persistence; [0129] b. Multivariate regression;
[0130] c. Autoregressive (AR) model; and [0131] d. Artificial
intelligence (AI).
[0132] 1.2.2 Persistence
[0133] The simplest statistical prediction is known as persistence
forecast. The persistence method d assumes that the conditions at
the time (t) of the forecast will not change, and the prediction is
set to equal to the last available measurement. In other words,
prediction is simply the last measured value:
{circumflex over (x)}.sub.t+1=x.sub.t (2)
[0134] The persistence method works well for a short period when
weather conditions change very little and features on the weather
maps move very slowly. However, if weather conditions change
significantly, the method usually breaks down and performs badly.
Although it is very simple, it is effective under certain
conditions.
[0135] 1.2.3 Multivariate Regression
[0136] A multivariate technique is the multiple linear
regression:
x ^ t + 1 = a 0 + i = 1 k a i y i ( 3 ) ##EQU00004##
[0137] where x is the prediction (wind speed or power output), y
are predictor variables, and k denotes the number of the predictor
variables. The predictor variables can include observations (such
as air pressure, temperature, humidity, etc.) at current and
previous times. Each of the predictor variables has its own
coefficient (a). All of the coefficients are derived from optimal
fitting (e.g., least squares) to the historical data. Forecasts
then are made by the equation. In some alternatives, Multivariate
regression is used.
[0138] 1.2.4 AR Model
[0139] A short term prediction technique is the more sophisticated
autoregressive or autoregression (AR) model. It takes into account
the fact that the atmospheric variables (like wind speed) have
"memory", and the model is defined as:
x ^ t + 1 = a 0 + i = 1 p x i ( 4 ) ##EQU00005##
[0140] where a are the autoregression coefficients, x are the time
series of wind speed, and p is the order (length) of the
autoregression. In summary, Equation (4) shows that the prediction
of wind speed or wind power can be estimated by a linear weighed
sum of previous observations. The weights are called the
autoregression coefficients. In some alternatives, AR models are
used.
[0141] 1.2.5 AI Model
[0142] AI attempts to simulate the human brain for searching,
cognition, learning, and reasoning. One AI technique is called
Artificial Neural Network (ANN). An example of an ANN model is
shown in FIG. 6. In the figure, the ANN consists of three layers:
input 610, hidden 620, and output 630. The input layer distributes
meteorological variables. The output layer are wind or wind power
predictions, and the hidden layer performs the mapping between the
input and output layers. N1, N2, and N3 are numbers of neurons at
the input, hidden, and output layers respectively. Once the ANN is
trained with historical data, it is expected to be able to
"remember" and, after comparing with current input, "pick up" the
right mapping relationship, and to make a projection into the
future by the relationship.
[0143] The ANN model is also used to predict directly power output
instead of winds. Although viable in certain circumstances, ANN is
good at pattern reorganization, not forecast. In some alternatives,
ANN is used.
[0144] 1.2.6 Physical Models
[0145] Physical models are based upon fundamental physical
principles of conservation of mass, momentum, and energy and the
equation of state for the atmosphere. These models consist of a set
of differential equations that are numerically solved on a
three-dimensional data grid with a finite resolution. In most (if
not all) cases, they are specially adapted to simulate the
atmosphere and involve various data assimilation schemes.
[0146] The physical approach can employ models of different scale
from synoptic scale, mesoscale, and microscale. Very often, a
single mesoscale model, taking forecasts from synoptic scale NWP
model as input, is used to predict wind power generation with
downscaling scheme. Among many numerical models, the Weather
Research & Forecasting (WRF) model stands out for its extensive
research efforts and popularity. WRF is a mesoscale model with a
capability of a fine resolution of 1 km. The numerical model output
can be either for the geographical point of the wind farm or for a
set of grids surrounding the farm.
[0147] In some cases, an ensemble prediction system is applied. The
system generates multiple realizations of weather variables by
using a range of different initial conditions for a numerical
model, or a range of different models. The frequency distribution
of the different realizations or models provides an estimate of the
density function, upon which probability forecasts are derived. An
ensemble forecast system requires huge computational resources, as
well as extensive knowledge and experience. For instance, the
European Centre for Medium-Range Weather Forecasts (ECMWF) produces
global weather forecasts with 51 ensemble members. Although
ensemble predictions are able to capture the dynamic change over
time in the density of a weather variable, they tend to
underestimate the spread of the density. For this reason, in some
alternatives, the ensemble forecasts need to be calibrated when
transforming to a forecast density.
[0148] 1.2.7 Hybrid Models and Model Output Statistics (MOS)
[0149] Both statistical and physical models have advantages and
disadvantages, and their performance varies with the forecast
horizon. In general, statistical models are simpler, cheaper, and
faster to develop. Also, they involve far less computing power and
resources. However, statistical models perform well only in a very
short term (0-3 hours ahead). Physical models outperform
statistical models after the forecast horizon.
[0150] In one embodiment, statistical and physical models are
combined together to become hybrid. Hybrid models use
state-of-the-art combination methods, and they often offer a
superior forecasting system.
[0151] In one embodiment, a family of Model Output Statistics (MOS)
methods is applicable in wind power forecasts. MOS is an objective
weather forecasting technique which consists of determining a
statistical relationship between the parameter being forecast and
values calculated by a numerical model. It is, in effect, the
determination of the "weather-related" statistics of a numerical
model.
[0152] 1.2.8 Model Applicability
[0153] Above are described various models that are used in various
embodiments of the systems and methods for wind prediction.
Applicability of types of forecast models is shown in Table 1. In
one embodiment, a 0-3 hour horizon, a statistical model
(particularly the AR model) combined with an ANN is used. Beyond
that horizon, physical models, plus MOS, are used.
TABLE-US-00001 TABLE 1 Advantages And Disadvantages Of Different
Models/Methods Model Characteristics Application Persistence
Simplest; Useful primarily in the benchmark to beat first hour
Multivariate Easy to develop Applicable to local area regression AR
model Sophistication; Less effective with effective nonstationary
and nonlinear problem ANN Able to tackle Questionable performance
nonlinear process; in making forecasts patter recognition Physical
models Dominated effec- Not so effective in the tiveness after
first a few hours; requires the first a few hours huge resources
and expertise Hybrid models Get merits from Needs to be trained
different models
[0154] 1.3 Prediction of WPREs
[0155] The prediction of WPREs differs from the general wind power
forecasts because of special features of the WPREs, i.e.,
abruptness and nonlinearity. It is these features that make WPRE
forecasting extremely difficult.
[0156] To make the difficult problem simpler, the prediction
process for WPREs can be decomposed into two actions: [0157] a.
Prediction of start and end of the ramp event (see red circles in
FIG. 7); and [0158] b. Prediction of power generation outside the
red circles 720.
[0159] In FIG. 7, the line 710 shows the wind power generation. The
methods and systems described herein allow for prediction within
the red circles 720 in FIG. 7. The focusing point should be on the
detection or prediction of start and end of such events.
[0160] 1.3.1 Challenges
[0161] The issues overcome by the embodiment of the model described
in Section 2 below are described herein. Predicting start and end
moments of a WPRE is crucial to the success of developing a WPRE
prediction system. However, such prediction is extremely
challenging because of the following issues.
[0162] (a) Model Incapability
[0163] Previous models/methods are of limited power to predict
WPREs and of limited success. Take an AR model as an example. It
has been proven a successful tool for nowcasting in general.
However, the AR model can only "see" and apply a hidden correlation
that exists in the time series. This means that the AR model is
unaware of a WPRE until the WRPE has become obvious for some time.
As a result, the AR model alone is unlikely to be able to predict
the start and end time of WPREs correctly. On the other hand, the
capability of a physical model is hampered by not only insufficient
understanding of underlying processes of WPREs, but also
computational limits (coarse resolution, long "spin-up" time,
etc.).
[0164] (b) Sensitivity of Turbine's Response to Winds
[0165] Prediction of a sudden change in wind speed has been
difficult. The response of turbines to winds at cut-in and cut-off
levels make prediction of a WPRE even more difficult. Winds
fluctuate frequently and swiftly. A small fluctuation around cut-in
or cut-off speed can cause a large gain or loss of wind power or a
ramp event.
[0166] Furthermore, once winds reach and move above a speed of
about 15 m/s, wind turbines begin to perform at their maximum
(100%) capability. This means that the relationship between wind
power output and wind speed is broken when wind speed is between 15
m/s and a cut-off speed (25 m/s). Therefore, a direct approach to
simulate the aggregated wind power time series must take this into
account.
[0167] (c) Cluster of Turbines
[0168] Wind turbines tend to be clustered in a small area for
economic reasons. The compacted cluster means that the wind farm is
highly sensitive to the change of winds on a local scale, while a
localized weather phenomenon (e.g., a local gust front) is harder
to observe and to predict.
2. WPRE Prediction
[0169] The methods and systems related to WPRE nowcasting take all
these factors into consideration and take a new lead. The new lead
comes from a simple but important fact; i.e., most, if not all,
weather systems leave "footprints" on the surface or near the
surface. Based on this fact, a ramp event can be observed or
detected by observations at surface levels and near-surface levels.
Therefore, in one alternative, a set of densely distributed surface
observation stations is used. A high spatial resolution off-site
observation network, including tower masts, a wind profiler, and an
Atmospheric Observation Network ((AON) see FIG. 8) serves this
purpose and should become an important part of the WPRE nowcasting
system.
[0170] One embodiment of a system for predicting WPRE is shown in
FIG. 9. The multi-model system consists of the following
components: [0171] a. A surface observation analyzer 910; [0172] b.
A vertical atmospheric analyzer 920; [0173] c. A mesoscale
numerical model 930; [0174] d. A neural network pattern recognizer
940; and [0175] e. A statistical forecast model 950.
[0176] As can be seen in FIG. 9, observation data from the AON and
Wind Production 960 is provided to all of the identifiers and
models: surface observation analyzer 910, vertical atmospheric
analyzer 920, mesoscale numerical model 930, neural network pattern
recognizer 940, and statistical forecast model 950. Analysis by the
surface observation analyzer 910 and vertical atmospheric analyzer
920 provide for the detecting of a WPRE signal in decision step
970. If a WPRE is identified, then the Neural Network 930 and
numerical model 940 estimate a magnitude and duration of the WPRE
in step 980. The result is fed into statistical model 950, and a
nowcast is produced in step 990. If no WPRE is detected in decision
step 970, then no estimate of the WPRE is made. Similarly, FIGS. 1,
2, 16a-e, describe similar embodiments and variations. More detail
on the specific embodiment of FIG. 9 is described below.
[0177] 2.1 Surface Observation Analyzer (SOA)
[0178] The aim of the analyzer, together with the Vertical
Atmospheric Analyzer, is to detect early signals of a ramp event,
based on fine spatial and temporal observation data around (or
off-site) and at (or on-site) the wind farm. The analyzer is based
on the fact that any weather system that causes a WPRE is
associated with the following features: [0179] a. significant
change in wind speed; [0180] b. strong vertical and horizontal wind
shears; [0181] c. pressure drop or surge at surface; [0182] d.
temperature increase or decrease; and [0183] e. shifts in
atmospheric stability.
[0184] Because of continuity of the atmosphere and motion of
weather systems, the above features, which can be observed, will
propagate along with the systems. Therefore, upwind observations of
the atmosphere give valuable information about the occurrence and
strength of a wind ramp event. This information, in turn, then is
used in short-term prediction of WPREs. The denser the observation
network, the better the information about the ramp events, and the
better the prediction of the events.
[0185] The AON is one of the data sources for the analyzer because
AON data: [0186] a. are measured by standardized
instruments/sensors; [0187] b. are frequently available--high
temporal resolution; [0188] c. are densely distributed in
space--high spatial resolution; [0189] d. have proven data quality
control schemes/processes; [0190] e. are reliable compared to other
kind of measurements; and [0191] f. are easy to process for real
time applications.
[0192] Based mainly on AON data, the SOA calculates spatial
gradients and temporal trends of: [0193] a. winds; [0194] b.
pressure; [0195] c. temperature; and [0196] d. humidity.
[0197] Results from calculation on the data (in some alternatives,
primarily up-wind observation data) provide information about the
existence and movement of synoptic scale and mesoscale weather
systems that may cause wind and wind power ramps.
[0198] Once a ramping signal is detected, propagation of the signal
(X) can be calculated (in a Cartesian coordinate) by:
v .fwdarw. .gradient. X = u .differential. X .differential. x + v
.differential. X .differential. y + w .differential. X
.differential. z ( 5 ) ##EQU00006##
[0199] This equation provides information to identify a WPRE and
detect its movement.
[0200] In one alternative, meteorological observations from surface
stations (AON stations and local climatological observation
stations) affect nowcasting algorithms/models in that: [0201] (1)
significant increase/decrease in wind speed at surface stations
accompany surge/drop of forecast and observed power output, with a
lead time. [0202] (2) The lead time varies, depending on distance
between the indicating station(s) and the wind farm, wind
direction, and average wind speed. [0203] (3) Variables measured at
the 10 m level, like wind speed and pressure gradient, are the most
significant predictors. [0204] (4) Meteorological variables
observed at multiple levels at AON 9 (see below) are important
predictors; these variables include vertical gradients of
temperature, wind shear, and Richardson number.
[0205] 2.2 AON Site Locations
[0206] The locations of AON sites in an exemplary system is
recorded in Table 2 and shown in FIG. 8. In this real world
example, the wind farm 824 is the area of interest. A ceilometer is
not included in the table or on FIG. 8, but is in operation in this
example.
TABLE-US-00002 TABLE 2 Lat Lon (decimal (decimal Site Tower
degrees) degrees) Legal County State AON 1 10 m 41.169423
-103.931288 SE4S23T14NR58W Kimball NE 802 AON 2 60 m 41.274166
-103.701385 NW4S13T15NR56W Kimball NE 804 AON 2 Wind 41.30195
-103.688033 NE4S1T15NR56W Kimball NE 806 Profiler AON 3 10 m
41.32095 -103.298233 NW4S33T16NR52W Cheyenne NE 808 AON 4 10 m
41.279311 -102.963733 NE4S17T15NR49W Cheyenne NE 810 AON 5 10 m
41.09091 -102.66823 NE4S23T13NR47W Cheyenne NE 812 AON 6 10 m
40.614366 -102.87415 SW4S32T8NR49W Logan CO 814 AON 7 10 m 40.6142
-103.569683 SE4S31T8NR55W Logan CO 816 AON 8 10 m 40.830183
-103.825916 NE4S23T10NR58W Weld CO 818 AON 9 60 m 41.075416
-103.60405 NW4S26T13NR55W Kimball NE 820 AON 10 10 m 41.018988
-103.337682 SW4S11T12NR53W Cheyenne NE 822
[0207] AON 2 804 is a special site in that the LAP-3000 wind
profiler is installed at a somewhat separated location from the AON
2 806. This is due to the need to separate the profiler from the 60
meter tower.
[0208] 2.2 AON Tower Instrumentation
[0209] 60 meter Towers: [0210] 60 meter height: Wind speed and
direction, Temperature, and Humidity Measurements. [0211] 30 meter
height: Wind speed and direction, Temperature, and Humidity
Measurements. [0212] 10 meter height: Wind speed and direction
Measurement. [0213] 2 meter height: Temperature and Humidity
Measurements. [0214] Pressure measurement instrument is in MAWS 301
cabinet.
[0215] 10 Meter Masts: [0216] 10 meter height: Wind speed and
direction Measurement. [0217] 2 meter height: Temperature and
Humidity Measurements. [0218] Pressure measurement instrument is
located in MAWS 110 cabinet.
[0219] 2.3 Vertical Atmospheric Analyzer (VAA)
[0220] Strong wind shear is often associated with severe synoptic
scale and mesoscale weather systems. Measurements of wind shear
thus provide important information about the possibility of WPREs.
One example of a VAA is the Vaisala LAP-3000. The Vaisala LAP-3000
Lower Atmosphere Wind Profiler is a Doppler radar which provides
vertical profiles of horizontal wind speed and direction, and
vertical wind velocity up to an altitude of 3 km above ground
level. It can also measure turbulence, boundary and mixing layer
heights, and combined with other data, atmospheric instability.
[0221] High vertical and temporal resolution data from the wind is
used to identify severe weather systems and estimate their
evolution (e.g., passage of a frontal system, its movement, and its
strength). Aided by surface wind measurements at AON, the wind
profiler data is used to interpolate wind profile to hub height at
which wind speed is critical for power generation.
[0222] Treating consecutive wind profiler measurements as a time
series, wind profiles extrapolate down the wind stream to a wind
farm and into the future.
[0223] FIG. 10 shows an example of the output from a Vertical
Atmospheric Analyzer. The Y-axis 1010 measures the height of the
wind in feet, and the X-axis 1020 measures the time. The intensity
of the wind is indicated by varying color, but is not shown in this
figure.
[0224] In one example, a LAP 3000 wind profiler is installed NW of
Logan-Peetz wind farm at AON 9 820, shown in FIG. 8. The AON09 wind
speed and the wind profiler <200 m averaged winds have a
correlation coefficient of .about.0.84, with wind profiler leading
on the average by 6 minutes.
[0225] 2.4 Mesoscale Numerical Model (MNM)
[0226] Performance of a mesoscale model largely depends on initial
and boundary conditions applied to the model. With data from
specifically designed AON, mast towers, wind profiler and other
sources, quantity and quality of initial and boundary conditions
improves significantly. This improvement enhances accuracy of the
model forecasts and subsequently leads to usefulness of numerical
model forecasts within a 3- or 4-hour horizon.
[0227] A fast four-dimensional data assimilation scheme is used to
take advantage of the sensor network. Once an early signal of a
(possible) ramp event is detected by SOA and VAA, the numerical
model is run to catch the signal and simulate the evolution of the
event. Results from the model run are used to aid the Neural
Network Pattern Recognizer and is integrated, together with the
Statistical Forecast Model, into the final step of the WPRE
prediction.
[0228] Besides fast and effective data assimilation (very likely to
be a data nudging) scheme, the mesoscale model is "light" and fast
in response. This requires good balance on the needs of a fine
spatial resolution and a fast execution time. In an alternative, a
daily or half-daily scan on mesoscale model outputs may still be
able to give us clues about severe weather systems and an early
alert.
[0229] 2.5 Neural Network Pattern Recognizer (NNPR)
[0230] The NNPR has to be trained, by feeding it teaching patterns
and letting it change its weights according to some learning rule,
before being put into real-time application. The variables as
inputs to the NNPR are mainly up-wind meteorological variables like
wind speed, wind direction, pressure, temperature, and
humidity.
[0231] Instead of training the neural network indiscriminately
based on all of the historical data, the NNPR uses carefully
selected data sets which involve WPREs only. Doing so increases the
effectiveness of the NNPR and, at the same time, reduces the burden
on computational resources.
[0232] 2.6 Statistical Forecast Model (SFM)
[0233] Statistical forecasting models can be categorized into two
models: [0234] a. Time series models (e.g., AR model), in which the
independent variable is time; and [0235] b. Explanatory models
(e.g., multivariate regression), in which the variable is one or
more factors.
[0236] Time series models assume that whatever forces have
influenced the variable(s) in question (such as wind power output)
in the recent past will continue into the near future. They are
very useful for short-term forecasting problems. The proposed SFM
is an AR model, although a multivariate regression model is not
ruled out completely.
[0237] The total actual power output from a wind farm is taken as a
time series, upon which statistical model(s) simulate. Doing so
bypasses the problem connected to the uncertainty of the power
curve as shown in FIG. 5. However, accurate prediction of wind
speed helps to adjust and improve direct forecasts on power
output.
[0238] Certain types of weather systems may take a certain route to
pass a wind farm. Nonhomogenous terrain has different influences on
wind fields and thus power output. Taking these into account, the
AR model is developed in predefined eight wind direction sectors,
each with 360/8=45 degrees. Wind-direction-dependent statistical
models are expected to perform better than a single general
model.
[0239] 2.7 Alternatives
[0240] 2.7.1 Radar Data
[0241] In one alternative to the method of FIG. 9, radar data is
included as part of the analysis. Weather radar is used to detect
weather systems and observe precipitation instantaneously. Modern
digital radar systems and data processors now have capabilities far
beyond early applications and are able to track storms fairly
robustly and accurately. This provides users of the system with the
ability to acquire detailed information of each storm, such as
location and movement. Naturally enough, radar and its data can be
thought of as a helpful tool that enhances the WPRE nowcasting
system. Precipitation events detectable by radar are tied to wind
events and durations based on the characteristics of the
precipitation event. Such radar data may also be used in relation
to the embodiments shown in FIGS. 1, 2, and 16a-e.
[0242] 2.7.2 Lagrangian Scalar Integration (LSI)
[0243] In one alternative, LSI is used to enhance detection of
synoptic and mesoscale features. LSI is a technique that is applied
to the gridded surface wind and scalar analyses. A grid of tracers
is specified over wind analysis at a resolution consistent with
features of interest (synoptic or mesoscale) and is advected
following the horizontal winds. Data are gathered along each
trajectory as a time series which is then time-averaged over some
fixed integration period. Specified features are then extrapolated
in line with advection. This technique is akin to releasing
"numerical weather balloons" and taking measurements along their
paths. Such a LSI may also be used in relation to the embodiments
shown in FIGS. 1, 2, and 16a-e.
[0244] 2.8. Summary
[0245] Among various models/methods, AR models and ANN methods
stand out for their power of dealing with time series and
recognizing patterns respectively.
[0246] Wind-power-ramp-related weather systems are associated with
strong wind shear within or above a boundary layer, shifts of
atmospheric stability, and more importantly, leave "footprints" on
the surface that can be observed. The surface may be the surface of
a weather system, a boundary layer, or other metrological
formation. In this disclosure, the phrase "wind farm" is a targeted
applicant of the nowcasting system. Once the concept, the logic,
the methods, and the system are proven, they can be applied to the
prediction of wind energy portfolios. In general, forecasts for
portfolios of wind farms are significantly more accurate than
forecasts for an individual wind farm, especially for large ramp
events.
3. Interface and Software Management System
[0247] 3.1 User Interface
[0248] In one embodiment, the systems and methods for wind
forecasting and/or WPRE prediction are implemented via a software
interface provided to the user of the forecasting. FIG. 11 shows an
example of the interface accessed by the user in a WPRE prediction
system. Here, the interface is a browser based system, which does
not require the user to install special software or hardware. From
this view of the interface, the user can see the projected and
historical wind power line 1110 and deviation band 1115, which
shows the expected range of wind power. On the left side of now
line 1116, historical data is shown; and on the right side of now
line 1116, predicted wind power data is shown. In this case, the
wind power monitored by the user is provided by two wind farms. The
power produced by the first wind farm is the area 1120, and the
power produced by the second is area 1130. These are automatically
totaled by the system. The user may also access the map of the wind
farms and view weather details overlaid on the map in area 1150. In
one alternative, this map is implemented in GIS, graphical
information system. Area 1140 denotes a WPRE, in this case a ramp
down predicted by the system.
[0249] Referring to FIG. 12, a ramp event summary table is shown
which is accessible to the user through the interface. The ramp
event summary tale provides important details to the user
concerning upcoming ramp events. The table identifies the wind farm
1210, the time until start 1220, the type of ramp 1230, the
magnitude of the ramp 1240 in MW, the start time of the WPRE 1250,
the end time 1260, and the duration 1270. In one alternative, this
information may automatically be exported to a power grid
management program to provide the user indications of what power
resources such as gas and coal systems should be ramped up or down,
or turned off or on.
[0250] FIG. 13 shows a window accessible through the interface by
the user. The user may set ranges for the magnitude of ramp events
and the rate of change for ramp events. For instance, in text entry
area 1310, the user may set a low point for the magnitude of
increase considered for a Level 1 ramp event. In text entry area
1320, the user may set a high point for the magnitude of increase
considered for a Level 1 ramp event. In text entry area 1330, the
user may set a low point for the rate of change considered for a
Level 1 ramp event. In text entry area 1340, the user may set a
high point for the rate of change considered for a Level 1 ramp
event. In alternatives, indicator colors are set, as well as other
alarms, such as audio alarms, email notifications, etc.
[0251] Some exemplary ramp event classification has been reset to
default values listed below:
[0252] L1 (Level 1) Magnitude: 25-100 MW
[0253] L2 Magnitude: 100-200 MW
[0254] L3 Magnitude: 200-300 MW
[0255] L4 Magnitude: >300 MW
[0256] Rate of Change (ROC): 0-5 MW/min, Color: Green
[0257] ROC: 5-8 MW/min, Color: Yellow
[0258] ROC: 8-10 MW/min, Color: Orange
[0259] ROC: >10 MW/min, Color: Red
[0260] Rate of change in MW/Minute
[0261] FIG. 14 shows one feature of the ramp window of FIG. 11.
When a user hovers over a ramp event 1410 with the mouse cursor,
details concerning the ramp event are displayed in popup 1420
including, but not limited to, the wind farms, the level of ramp
event, the start and end times, and the duration. FIG. 15 shows
another feature of the ramp window of FIG. 11. The area on the
window 1510 denotes three distinct time periods for predictions,
from t=0 to t=60 time period 1520, from t=60 to t=120 time period
1530, and from t=120 to t=180 time period 1540. The general
volatility of these time periods is displayed to the user by color
coding time periods 1520, 1530, 1540, wherein red indicates high
volatility, yellow medium, and orange low.
[0262] FIG. 18 shows one embodiment of a ramp window and features
of a ramp window. The ramp window includes the sliding window 1810
concept with a wind of a fixed size based on a ratio of fixed
change in power over fixed change in time. In short, this means
that each category of ramp event is defined as described above in
relation to ROC. Here the change in power 1830 divided by the
change in time 1820 yields an orange ROC event 1840. Note how even
though the duration of the ramp event, time 1820 is small, it still
is a ramp event based on the rate of change and its qualification
as a sufficient change in power 1830 to qualify as a level of ramp
event. In event 1850, even though the rate of change is fast, the
duration of the event and the size of the change of power are
insufficient to qualify as a ramp event. Event 1860 is a false ramp
event since duration and magnitude are not sufficient. Even though
a ramp entry and exit are discovered, insufficient magnitude and
duration exist to qualify as a ramp event. Another down ramp with a
magnitude 1875 and duration 1870 and another up ramp with a
magnitude 1885 and duration 1880 are also shown. The down ramp
qualifies as an orange event based on its ROC and the up ramp as a
red event based on its rate of change. The scale on y axis 1895 is
MW and x axis 1890 is 45 minute intervals.
[0263] 3.1 Architecture
[0264] An example of the architecture for one embodiment of a
system for nowcasting wind events is shown in FIG. 17. The
embodiments shown in other figures such as 1, 2, 9, and 16a-e may
be implemented using similar architectures. The architecture
includes a WENDSS system 1710. The WENDSS system is a single server
that contains a database 1711, services (SOA, messaging, UI layer,
Bus 1715, etc.), an AA forecast algorithm layer 1712, and
interfaces to external systems and databases. The WENDSS internal
database stores all system data including AON observations,
historic PI data, and time-series forecasts 1714 created by the AA
Engine 1712. AON and PI data 1713 are provided to the AA Engine
1712 for analysis. In alternatives, the WENDSS need not be a single
server, and instead may be a distributed system, between two or
more computers. The WENDSS may be a cloud computing system, or
implemented in a variety of hardware and software configurations.
XML observation data 1771, XML Time Series Data 1772, and Formatted
Data files 1773 are all examples of data that are provide to WENDSS
1710.
[0265] The architecture includes an AON 1720. AON 1720 is a
customized environmental data collection network that surrounds the
wind farm. The network is made up of multiple AON stations. Most
AON stations collect wind speed, wind direction, RH, and
temperature data. In alternatives AON stations have wind profilers
and/or ceilometers. An example of an AON is shown in FIG. 8
above.
[0266] NGMetman 1730 is a shared environmental database. It is used
to ingest the AON data. NGMetman output jobs, specific for the
WENDSS system, allow data files to be delivered to the WENDSS
system on a regular interval. The AON data is used by the AA engine
1712 to produce the power forecast. In one alternative, the
NGMetman 1730 produces XML data for the WENDSS system 1710.
[0267] PI Server 1740 provides plant information from a power
company. The PI system (Plant Information System) collects data
related to a wind farm including individual turbine and aggregate
power generation data. The server in diagram receives PI data from
the main PI server operated by the power company, and then forwards
this data to the WENDSS system on a regular interval.
[0268] NWS data 1750 represents National Weather Service public
data sources that also are used by the AA. Communication is
provided between the WENDSS 1710 and PI Server 1740 and WENDSS 1710
and NGMetman 1720 via Web Service Interface to ESB 1774 and 1775.
Communication is provided between the WENDSS 1710 and NWS data 1750
via FTP or LDM Interface to ESB 1776.
[0269] WENDSS User Interface 1760 is web-based user interface that
presents system data to the user. Users access the UI using their
existing network and associated PCs. Some details of one possible
UI are described above. Numerous alternatives to the architecture
described herein will occur to those skilled in the art in light of
the disclosure of this architecture and the disclosure of systems
and methods herein.
4. Examples
[0270] Some examples of how the systems and method described herein
can be used are as follows.
The Customer Problem
[0271] Tom is an Operations Analyst in Xcel Energy's Real-Time
Commercial Operations group that is responsible for dispatching and
committing generation to meet their customers' energy needs while
maintaining reliability on their service territory grid. His duties
are focused upon Xcel Energy's Public Services of Colorado (PSCo)
service territory that serves 1.35 million customers with an
average load (electricity usage) ranging from 3,000 to 6,000 MW.
Tom's job has become much more difficult in the past year as Xcel
Energy continues to establish power purchase contracts with wind
farms to meet the State of Colorado's Renewable Portfolio Standard
that mandates 15% of their power delivered by 2015 must come from
renewable sources. PSCo now integrates up to 1,258 MW of wind
energy into their grid, which means under ideal wind conditions, up
to 35% of PSCo's electricity delivered is coming from wind farms.
Tom is supportive of Xcel Energy's adoption of wind energy, but
there are a number of problems it creates for his day-to-day
duties.
[0272] Xcel Energy's contracts mandate that they must pay each wind
farm owner/operator for the power they produce and feed into PSCo's
grid. Each wind turbine is rated upon a power curve that shows how
much power is to be generated given the wind speed. Most modern
wind turbines begin to generate power once winds exceed 4 ms-1 and
quickly increase the amount of power generated to the full rated
capacity of the turbine when winds exceed 12 ms-1. Wind speed
fluctuations in this 4-to-12 ms-1 range cause rapid changes in
energy production that Tom and his colleagues have never
experienced with the coal and combined cycle gas plants they are
accustomed to. They refer to these rapid changes in wind energy
generation as "ramp events", due to the sharp curve typically seen
in their total generation portfolio output. Ramp events make Tom's
job difficult because he must keep PSCo's generation within their
Area Control Error (ACE) of +/-56 MW more than 90% of the time (as
measured in 10 min windows each month), which is mandated by the
North American Electric Reliability Corporation (NERC) to ensure
the transmission grid continues to operate properly. ACE is the
instantaneous difference between net actual and scheduled
interchange, taking into account the effects of frequency bias
including a correction for meter error between PSCo and its
neighboring NERC Balancing Areas. Additionally, PSCo's energy
planning and trading groups cannot provide Tom with the best mix of
generators to optimize their financial performance when the wind
energy forecasts are not reliable. Tom realizes that the problems
are only going to get worse as Xcel Energy continues to add more
wind energy to its system.
Use Case #1 Overnight Ramp-Up Event (Operations Analyst
Perspective)
[0273] Tom is an Operations Analyst in Xcel Energy's PSCo Real-Time
Commercial Operations group on 24 Oct. 2008 working the late night
shift at 2:00 am. The load forecast provided by Xcel Energy's
meteorologist is typical of this time of year, as load will drop
off into the early morning hours as people are sleeping and the
temperatures cool. However, in his routine check the Vaisala WENDSS
is forecasting a "Yellow Level 3 Advisory" Ramp-Up event to
initiate between 3:40 and 4:00 am. WENDSS is an exemplary
embodiment of systems and methods described herein. Wind energy
generation is currently near 300 MW, so Tom realizes the ramp event
will bring the contribution from wind to near its full rated
capacity. Although the ramp rate will not be at a dangerous level,
the timing of this ramp event greatly concerns Tom because it will
coincide with the lowest loads of the day near 3,100 MW. Xcel
Energy PSCo can only ramp down their coal generators to an
aggregate 2,500 MW system minimum overnight to ensure reliable next
day operation. Given the Vaisala WENDSS advisory, Tom recognizes
that their total generation (base load+wind) will exceed the PSCo
Balancing Area load. Tom immediately contacts the trading desk to
initiate their planning efforts to offload generation to
neighboring utilities. The trading desk is able to offload some
energy starting at 4:00 am after the ramp initiates, but Tom has no
other choice but to contact the Logan-Peetz Wind Farm at 4:15 am
and ask them to open the breaker to their grid interconnection. Tom
continues to track the Vaisala WENDSS short-term wind generation
forecast and is able to bring Logan-Peetz back on the grid at 6:10
am as PSCo Balancing Area load increases to levels sufficient to
support minimum base load and the full capacity wind generation as
people wake-up and turn their lights on. The Vaisala WENDSS'
accurate forecast of the timing and magnitude of the ramp event
allowed Xcel Energy to stay within their ACE 80% of the time and
handle the early morning ramp event in the most cost effective
manner.
Use Case #2 Late Afternoon Ramp-Up Event (Operations Analyst
Perspective)
[0274] Tom is working the daytime shift as an Operations Analyst in
Xcel Energy's PSCo Real-Time Commercial Operations group on 11 Aug.
2008. It is a hot summer afternoon in Colorado and customer load
continues to rise to near 5,800 MW as energy consumption rates near
their peak for the day. Xcel Energy's meteorologist advised Tom's
team earlier in the morning that the temperatures would be very hot
in the upper 80's and the winds would be light, with a chance for
isolated thunderstorms. To meet this anticipated demand, Xcel
Energy scheduled to bring two combined cycle gas plants online in
the late afternoon since there would be little available wind
energy.
[0275] At 2:00 pm, Tom felt he had everything under control and he
continued to monitor his Energy Management System display that
tracks the base load generators and the Automated Generation
Control (AGC) that slowly increases generation from their coal and
combined cycle gas plants to meet the PSCo Balancing Area load and
maintain their ACE compliance. As expected, wind energy generation
had been less than 100 MW for most of the day, but a flashing alert
on the Vaisala Wind Energy Decision Support System (WENDSS) caught
his eye. Tom acknowledged the alert and saw a Red Level had been
issued to occur between 3:05 and 3:15 pm due to strong outflow
winds from thunderstorms in rural northeast Colorado, where most of
their wind farms are clustered. If the event was as strong as
forecasted, more than 20 MW of wind energy would be coming online
every minute, making it difficult to stay in ACE compliance. Tom
immediately began following the safety procedure developed by his
manager. He calls the their largest combined cycle gas plant to
initiate the 10 minute process necessary to obtain manual control
of their system. Tom could rely upon using their "go-to-minimum"
command they can transmit to each of their plants, but this only
allows for a 3 MW min-1 ramp down for each plant and would not be
sufficient to balance the grid due to the extreme wind event. At
3:10 pm, Tom monitors the Vaisala WENDSS and sees the wind energy
generation begin to ramp up from 80 MW to 250 MW within 20 minutes
on the real-time display. He is able to manually instruct the
combined cycle gas plant operator to cycle down from 700 MW to 400
MW to keep the ACE within its boundaries thanks to his advanced
planning. The Vaisala WENDSS 3-hour forecast shows the ramp will
continue for another hour, but after the thunderstorms pass the
load will begin to drop as the wind power generation drops, so Tom
will be able to relinquish control of the combined cycle gas plant
back to the operators and pass along a normal workload to the next
shift. Thanks to the Vaisala WENDSS, Tom saved Xcel Energy up to
$50,000 it would have been obligated to pay their wind farms if
generation had been curtailed.
Use Case #3 Late Afternoon Ramp-Up Event (Trader Perspective)
[0276] Karl is working the same shift as Tom, as a Trader in Xcel
Energy's PSCo Real-Time Commercial Operations group on 11 Aug.
2008. It is his responsibility to optimize Xcel Energy's financial
performance by selling energy at the highest price when generation
exceeds load, and purchase energy at the lowest price when load
exceeds generation. Xcel Energy was not expecting much wind
contribution to their generation mix, so their three largest coal
generation plants and combined cycle gas unit were scheduled for
significant operation during the PSCo Balancing Area peak load in
the late afternoon. However, the isolated thunderstorm activity in
the vicinity of Xcel Energy's wind farms introduced unexpected
levels of wind energy generation that created an ACE of +75 MW at
5:55 pm just after the wind ramp event initiated. Karl had
consulted the WENDSS as part of his standard procedures 15 minutes
ago so he was prepared to assist Tom by offloading some of their
combined cycle gas generation to a neighboring Balancing Area that
was struggling to cope with the warm temperatures and peak load
conditions. Karl was able to sell electricity at a significant
price premium due to Xcel Energy's advantage of having wind energy
on their system. The Vaisala WENDSS' accurate prediction of the
ramp event with at least 30 minutes lead time led to improved
reliability and financial gain under difficult conditions.
Use Case #4 Late Afternoon Ramp-Down Event (Operations Analyst
Perspective)
[0277] Tom is an Operations Analyst in Xcel Energy's PSCo Real-Time
Commercial Operations group on 15 Jun. 2009 at 3:30 pm. During his
routine checks, he notices a flashing red notification on the
Vaisala WENDSS notifying him of an "Orange Level 4 Alert" Ramp-Down
to initiate between 5:00 and 5:20 pm. Tom has come to recognize
that it is difficult to pinpoint the timing of ramp events, but the
uncertainty timing window provided by the Vaisala WENDSS allows him
to initiate the necessary actions in response to the suspected
event. The day-ahead generation schedule issued by Xcel Energy
limited the amount of base load on-line, as wind energy was
expected to be a significant contributor through the day. Tom
continued to closely watch the WENDSS and checked it every five
minutes as the forecast was updated using the latest information
from the atmospheric observation networks surrounding the wind
farms. He appreciates the forecast probability information in the
WENDSS graphical display. The confidence bands around the forecast
allow Tom to run various simulations in his GenTrader application
running on another display that provides decision support regarding
the most economical generators to use when reserves must be brought
online.
[0278] In a situation where the PSCo Balancing Area's ACE drops
quickly and approaches the negative -56 MW, Tom's best available
option is to dispatch the appropriate number of natural gas peaker
units with at least 10 minutes lead time and have them prepared as
spinning reserves before the wind ramp-down initiates. At 4:55 pm
the Vaisala WENDSS forecasts the ramp down will be moderate, with a
peak rate of change of -16 MW min-1 with an expected duration of 70
minutes. The Vaisala WENDSS is having a difficult time resolving
the magnitude of the event, as it could range from -250 to -425 MW
according to the confidence bands. Based on this information, Tom
dispatches 8 natural gas peaker units and advises the trading group
to move forward with purchasing electricity on the open market as a
necessary precaution. The short duration of the ramp will eliminate
the need to dispatch operating reserves, as the current base load
generators can be cycled up to appropriate levels while the ramp
event is ongoing.
[0279] At 5:20 pm the ramp-down event initiates and Tom commits the
spinning reserves in an efficient manner maintaining the ACE at a
comfortable -30 MW level. Tom continues monitoring the WENDSS at
5:40 pm as wind generation has dropped 180 MW to an aggregate of
200 MW across PSCo. Thankfully the peak ramp rate has only reached
-11 MW min-1, but Tom is cautious since the wind generation
volatility index is at a Red level, indicating a majority of the
wind turbines are operating at wind speeds between 4-to-12 ms-1 and
could change power output very quickly. The WENDSS indicates that
the ramp rate should level out at 6:30 pm, so Tom issues requests
to de-commit the natural gas peakers one-by-one as the base load
generators have been cycled up to levels to support the diminishing
load as it drops below 4,000 MW into the nighttime hours.
[0280] In the post-event analysis, Tom's manager complimented him
on his use of the Vaisala WENDSS during the event to support his
decision-making process. The early notification of the potential
ramp-down allowed Tom to stage the reserve units well in advance to
ensure reliability of the grid. In fact, Tom was able to dispatch
lower cost flex reserve units that need 20-40 minutes lead time
thanks to the WENDSS. Tom's manager acknowledged that Xcel Energy
had incurred $100,000 in expenses tied to dispatch and utilization
of the reserve units, but it ensured they did not have to shed load
and revenue by opening breakers to large industrial customers. Even
more importantly, Xcel Energy maintained the trust and confidence
from their customer base that they can keep the lights on under the
most difficult of circumstances at competitive market prices.
[0281] In all cases of the above-described embodiments, the results
of any of the transformations of data described may be realized by
outputting the results by transforming any physical or electronic
medium available into another state or thing. Such output includes,
but is not limited to, producing hardcopy (paper), sounds, visual
display (as in the case of monitors, projectors, etc.), tactile
display, changes in electronic medium, etc. The foregoing
description of the embodiments of the inventions has been presented
only for the purpose of illustration and description and is not
intended to be exhaustive or to limit the inventions to the precise
forms disclosed. Numerous modifications and adaptations are
apparent to those skilled in the art without departing from the
spirit and scope of the inventions.
* * * * *